Environment: dmd 2.052, 64-bit Ubuntu.

The following program passes the two asserts, showing that the array elements are not initialized as NANs. The dumpBytes() function shows that the bit pattern is different than a local variable's:

import std.stdio;

void main()
{
    double[1] a;
    auto myNaN = double.nan;

    assert (a[0] != myNaN);        // <-- passes; BUG
    assert (a[0] != double.nan);   // <-- passes; BUG

    dumpBytes(a[0]);
    dumpBytes(myNaN);
}

void dumpBytes(T)(ref T var)
{
    const ubyte * beg = cast(ubyte*)&var;

    foreach (bayt; beg .. beg + T.sizeof) {
        writef("%02x ", *bayt);
    }

    writeln();
}

The bug manifests itself for float and real as well and the bit pattern changes depending on whether -m32 or -m64 is used. Here is one output of the program with -m64:

00 00 00 00 00 00 fc 7f 
00 00 00 00 00 00 f8 7f 

Ali